Material design device and material design method

Abstract

Provided are a material design device and a material design method having a high degree of freedom such that a plurality of materials can easily be added. A material design device 1 according to an embodiment comprises a control unit 10, a storage unit 20, an input unit 30, and an output unit 40. The control unit 10 is provided with an embedding processing unit 11, a first linear processing unit 12, a combining unit 13, a generation unit 14, a separation unit 15, a classification processing unit 16, and a second linear processing unit 17.

Description

Material Design Device and Material Design Method

[0001] The present invention relates to a material design device and a material design method.

[0002] Conventional material design has been carried out based on experiments and has been carried out by inheriting the past experiences of material designers. Past empirical rules serve as judgment materials for material design. On the other hand, when empirical rules and the like are not left as information such as data, the accurate properties of materials often become uncertain, and experiments may need to be redone. Also, even when the type and amount of materials change slightly, it is necessary to repeat measures such as mixing the materials and conducting experiments, which has required a lot of effort from the economic and time perspectives.

[0003] When mathematically processing the data obtained from experiments, usually a large number of data are required, but due to economic and time constraints of conducting experiments multiple times, a lot of effort has been required.

[0004] Therefore, in recent years, in the design of new materials, for example, techniques for predicting material properties using AI (Artificial Intelligence) technologies such as machine learning or deep learning have been disclosed (for example, Patent Document 1, etc.). By using AI technologies, the number of experiments can be reduced, and it is expected to improve the development efficiency of new materials.

[0005] Japanese Unexamined Patent Application Publication No. 2023 - 161193

[0006] Conventional material design has been carried out by experimental design methods, etc., calculating the interaction between materials, etc., and making inferences from the results obtained there. This method has problems such as being complex, requiring input of detailed values other than the amount of materials using spreadsheet calculations, etc., and the complexity increasing as the number of materials increases.

[0007] Furthermore, as in Patent Document 1, when designing materials using machine learning or deep learning, a method is used in which the quantity and / or classification parameters of a substance are input to a predetermined statistical model (including a neural network). However, while this method simplifies the model, it has the drawback of being usable only within a specific range of materials or a specific range of material performance. Also, if the goal is simply to implement the results of mathematical processing, it is possible to do so using only machine learning or deep learning. On the other hand, when focusing on the human interface (the part related to input and output), there is a problem that the complexity of the method makes it very difficult to use for designers other than data scientists in the field of materials design, or for design beginners.

[0008] Therefore, in order to solve these problems, the present invention aims to provide a highly flexible material design apparatus and material design method that allows for the easy addition of multiple materials.

[0009] To achieve this objective, the material design apparatus and material design method of the present invention are characterized by having the following inventive features.

[0010] The material design apparatus is a material design apparatus for designing a material containing multiple substances, comprising: an input unit for inputting the substance names of the multiple substances and the weight portion, which is the amount of the substance corresponding to the substance name relative to the total mass; an embedding processing unit for the multiple substance names to generate a substance name vector; a first linear processing unit for the weight portion corresponding to the multiple substance names to generate a weight portion vector; a coupling unit for superimposing the substance vector and weight portion vector corresponding to the multiple substances to generate a first coupling vector; a generation unit for inputting the first coupling vector into a trained model to generate a second coupling vector by superimposing a characteristic name vector representing the characteristic name of at least one characteristic of the material to be designed and a characteristic quantity vector representing the quantity of at least one characteristic; a separation unit for separating the superposition of the second coupling vector; a classification processing unit for at least one characteristic name vector to generate a characteristic name; and a second linear processing unit for at least one characteristic quantity vector to generate a characteristic quantity. The system is characterized by comprising an output unit that outputs the characteristic name and characteristic quantity of at least one or more characteristics.

[0011] Furthermore, the material design apparatus is a material design apparatus for designing a material containing multiple substances, comprising: an input unit for inputting characteristic names and quantities corresponding to at least one characteristic; an embedding processing unit for performing embedding processing on at least one characteristic name and generating a characteristic name vector; a first linear processing unit for performing linear processing on quantities corresponding to at least one characteristic name and generating a characteristic quantity vector; a coupling unit for superimposing the characteristic name vector and characteristic quantity vector corresponding to at least one characteristic and generating a first coupling vector; a generation unit for inputting the first coupling vector into a trained model and generating a second coupling vector by superimposing a substance name vector representing the substance names of multiple substances constituting the material to be designed and a weight portion vector representing the weight portion of the multiple substances as a whole; a separation unit for separating the superposition of the second coupling vector; a classification processing unit for performing classification processing on multiple substance name vectors and generating substance names; and a second linear processing unit for performing linear processing on multiple weight portion vectors and generating weight portions. It is characterized by having an output unit that outputs the names of multiple substances and their weights.

[0012] In any of the material design apparatuses described above, the trained model is preferably a model based on an encoder-decoder architecture having an encoder and a decoder, a recurrent neural network (RNN) model, or a transformer model.

[0013] The material design method is a material design method for designing a material containing multiple substances, comprising: an input step of inputting the substance names of the multiple substances and the parts by weight which are the amounts of the substances corresponding to the substance names relative to the total mass; an embedding step of performing embedding processing on the multiple substance names to generate substance name vectors; a first linear processing step of performing linear processing on the parts by weight corresponding to the multiple substance names to generate part by weight vectors; a coupling step of superimposing the substance vectors and part by weight vectors corresponding to the multiple substances to generate a first coupled vector; a generation step of inputting the first coupled vector into a trained model to generate a second coupled vector by superimposing a characteristic name vector representing the characteristic name of at least one characteristic of the material to be designed and a characteristic quantity vector representing the quantity of at least one characteristic; a separation step of separating the superposition of the second coupled vector; a classification processing step of performing classification processing on at least one characteristic name vector to generate characteristic names; and a second linear processing step of performing linear processing on at least one characteristic quantity vector to generate characteristic quantities. The invention is characterized by comprising an output step that outputs the characteristic name and characteristic quantity of at least one or more characteristics.

[0014] Furthermore, the material design method is a material design method for designing a material containing multiple substances, comprising: an input step of inputting a characteristic name and a quantity corresponding to at least one characteristic; an embedding step of performing embedding processing for at least one characteristic name to generate a characteristic name vector; a first linear processing step of performing linear processing for the quantity corresponding to at least one characteristic name to generate a characteristic quantity vector; a coupling step of superimposing the characteristic name vector and the characteristic quantity vector corresponding to at least one characteristic to generate a first coupled vector; a generation step of inputting the first coupled vector into a trained model to generate a second coupled vector by superimposing a substance name vector representing the substance name of each of the multiple substances constituting the material to be designed, and a weight vector representing the weight portion of the multiple substances in total; a separation step of separating the superposition of the second coupled vector; a classification processing step of performing classification processing on the multiple substance name vectors to generate substance names; and a second linear processing step of performing linear processing on each of the multiple weight portion vectors to generate their respective weight portions. It is characterized by comprising an output process that outputs the names of multiple substances and the corresponding parts by weight.

[0015] According to the material design apparatus and material design method of this configuration, the name of a substance or property can be treated as a "token" using "sequence to sequence," one of the neural network configurations, and the quantity corresponding to that name of substance or property can be treated as a "regression." In this way, by treating the name of a substance or property as a "token" as the "language" of natural language processing, it becomes possible to handle any name of a substance or property. Furthermore, since the quantity of a substance or property can be input as a "regression," it becomes possible to handle it linearly.

[0016] Furthermore, if the material design device and method have learned the types and quantities of substances, it becomes possible to calculate characteristic values ​​as inference results simply by inputting the name and quantity of a specific substance. Conversely, if the material design device and method have learned the types and quantities of properties, it becomes possible to calculate the composition and structure of a substance as inference results simply by inputting the name and quantity of a specific property. This allows for the easy addition of multiple substances or properties, increasing the degree of flexibility.

[0017] This is a block diagram showing the configuration of a materials design apparatus according to one embodiment of the present invention. This is a flowchart showing the control process of a materials design method performed by a materials design apparatus according to one embodiment of the present invention. This is a diagram showing the learning curve of a materials design apparatus according to the first embodiment. This is a diagram showing the learning curve of a materials design apparatus according to the second embodiment. This is a diagram showing the learning curve of a materials design apparatus according to the third embodiment. This is a diagram showing the comparison result between the inference result of the materials design apparatus in the third embodiment and the training data. This is a diagram showing the learning curve related to data group A of the materials design apparatus according to the fourth embodiment. This is a diagram showing the learning curve related to data group B of the materials design apparatus according to the fourth embodiment. This is a diagram showing the learning curve of the result of transfer learning of the results learned with data group A using data group B. This is a diagram showing the learning curve of the result of transfer learning of the results learned with data group B using data group A. This is a diagram showing the comparison result between the inference result of a materials design apparatus learned based on data group A and the training data. This is a diagram showing the comparison result between the inference result of a materials design apparatus learned based on data group B and the training data. This is a diagram showing the comparison result between the inference result of a materials design apparatus that transfers data group A based on data group B and the training data. This is a diagram showing the comparison result between the inference result of a materials design apparatus that transfers data group B based on data group A and the training data. This is a diagram showing the learning curve related to a materials design apparatus according to the fifth embodiment. This figure shows the learning curve of the material design device 1 when no information identifying the bearing type is input in the fifth embodiment. This figure shows the comparison result between the inference result of the material design device 1 according to the fifth embodiment and the learning data. This figure shows the comparison result between the inference result of the material design device and the learning data when no information identifying the bearing type is input in the fifth embodiment. This figure shows the learning curve for the material design device according to the sixth embodiment. This figure shows the verification curve for the material design device according to the sixth embodiment. This figure shows the comparison result between the inference result of the material design device according to the sixth embodiment and the learning data. This figure shows the estimated wear amount in the sixth embodiment. This figure shows the estimated friction coefficient in the sixth embodiment. This figure shows the estimated wear amount in the sixth embodiment. This figure shows the estimated friction coefficient in the sixth embodiment. This figure shows the estimated wear amount and friction coefficient in the sixth embodiment.This figure shows the estimated friction coefficient results from the sixth embodiment.

[0018] (Configuration) A material design apparatus 1 according to one embodiment of the present invention will be described in the following description and drawings. In this embodiment, information analysis processing as a predetermined process will be described as an example. As shown in Figure 1, the material design apparatus 1 of this embodiment comprises a control unit 10, a storage unit 20, an input unit 30, and an output unit 40.

[0019] Each of the control unit 10 and the storage unit 20 consists of an arithmetic processing unit such as a CPU, or memory, an I / O interface, and a storage device such as ROM, RAM, or HDD. The arithmetic processing unit consists of one or more CPUs that read necessary software and data from memory and execute specified arithmetic processing on said data according to the software, and, if necessary, communication equipment, storage device (said memory), etc.

[0020] The control unit 10 comprises an Embedding processing unit 11, a first Linear processing unit 12, a coupling unit 13, a generation unit 14, a separation unit 15, a classification processing unit 16, and a second Linear processing unit 17. Each of the Embedding processing unit 11, the first Linear processing unit 12, the coupling unit 13, the generation unit 14, the separation unit 15, the classification processing unit 16, and the second Linear processing unit 17 that constitute the control unit 10 is composed of an arithmetic processing unit such as a CPU, or memory, an I / O interface, and a storage device such as ROM, RAM, or HDD.

[0021] The Embedding Processing Unit 11 converts the words (tokens) input to the Input Unit 30 into a vector representation, as will be described later. In other words, the Embedding Processing Unit 11 performs mathematical embedding of the words into a vector space to make them suitable for machine learning such as natural language processing, which will be described later.

[0022] As will be described later, the first Linear processing unit 12 performs Linear processing on the real number input to the input unit 30. As will be described later, from the viewpoint that the vector generated by the Embedding processing unit 11 and the vector generated by the first Linear processing unit 12 are combined, it is preferable that the dimension of the vector generated by the first Linear processing unit 12 is the same as the dimension of the vector generated by the Embedding processing unit 11.

[0023] The coupling unit 13 superimposes the vector generated by the Embedding processing unit 11 and the vector generated by the first Linear processing unit 12. In other words, the coupling unit 13 generates a first combined vector as a result of the superimposition.

[0024] The generation unit 14 generates a second connected vector as a result of the first connected vector via a trained model stored in the storage unit 20, as described later. When multiple first connected vectors are input to the generation unit 14, the multiple first connected vectors may be combined in the generation unit 14 or the connecting unit 13 before being input to the generation unit 14. The trained model input to the generation unit 14 is preferably a model with an encoder-decoder architecture having an encoder and decoder as a recurrent model that can perform sequence-to-sequence processing, a recurrent neural network (RNN) model, or a transformer model, from the viewpoint of treating the vector generated by the Embedding processing unit 11 and the vector generated by the first Linear processing unit 12 as sequential data.

[0025] The generation unit 14 may include a transfer learning unit. That is, the trained model may be generated by transfer learning an existing trained model. The method of transfer learning is not particularly limited, and existing methods can be used.

[0026] The dimensions of the first and second bond vectors may be the same or different. Furthermore, even if multiple first bond vectors are input to the generation unit 14, the number of first and second bond vectors may be the same or different. If the vector generated by the generation unit 14 is composed of multiple second bond vectors, the generation unit 14 or the separation unit 15 separates them into their respective second bond vectors.

[0027] The separation unit 15 separates the superposition of the second bond vectors, as will be described later. As a result, multiple vectors are generated as a result of separating the second bond vectors.

[0028] As described later, the classification processing unit 16 performs a classification process on one of the multiple vectors resulting from separating the second combined vector. Through the classification process, the classification processing unit 16 classifies the single vector into a word (token), etc.

[0029] The second Linear processing unit 17 performs Linear processing on the vectors other than one of the multiple vectors obtained as a result of separating the second bond vector. Note that the Linear processing performed by the first Linear processing unit 12 and the Linear processing performed by the second Linear processing unit 17 are usually different, but they may be the same.

[0030] The storage unit 20 is configured to store and hold the results of calculations performed by the arithmetic processing unit.

[0031] The arithmetic processing unit consists of an information processing unit (CPU) that reads software and data from a designated area of ​​memory constituting the storage device as needed, and then performs a specified arithmetic processing on that data in accordance with the software, and communication equipment, storage device (the memory), etc., as needed.

[0032] The input unit 30 consists of operation buttons and a microphone, enabling user-operated touch-based operation or non-contact operation via voice commands. The output unit 40 consists of a display device and an audio output device (speaker), which displays the design results calculated by the material design device 1 as visual information or outputs them as auditory information such as audio content. The input unit 30 and the output unit 40 may also consist of touch panel displays.

[0033] (Function) Next, a material design method in a material design apparatus 1 according to one embodiment of the present invention will be described in the following description and drawings. Figure 2 is a flowchart showing the control process of the material design method performed by the material design apparatus 1 according to one embodiment of the present invention. In this embodiment, a material design apparatus 1 is described that, when the names of multiple substances and the weight portion, which is the amount of the total mass of the substances corresponding to the names of the substances, are input, outputs the characteristic name and characteristic quantity of at least one or more characteristics.

[0034] First, material name data, representing the names of multiple substances, and part-of-weight data, representing the amount of each substance relative to the total mass, are transmitted to or uploaded to the material design device 1 via the input unit 30 (Figure 2 / STEP 1).

[0035] Here, "substance" refers to the material that makes up a material, and is a concept that includes elements and compounds. Examples of substance name data in STEP 1 include metals or special metals such as aluminum and iron, organic substances such as carbon and polypropylene, or engineering plastics. Alternatively, the labels used for classification may be entered as substance name data after prior classification. That is, when aluminum is selected as the substance, [1] as the label for aluminum may be entered as substance name data.

[0036] Parts by weight data is a quantity that represents the mass of the substance corresponding to the substance name data relative to the total mass of the material. Note that the quantity representing the amount of the substance corresponding to the substance name data relative to the total mass of the material is not limited to weight ratio; values ​​such as composition ratio or volume ratio may also be used.

[0037] In material design, materials are typically composed of multiple substances. Therefore, in STEP 1, substance name data and parts-of-weight data are entered for each of these substances.

[0038] In STEP 1, substance name data and parts-of-weight data may be confused. Therefore, it is preferable that the input be in a manner that prevents confusion between substance name data and parts-of-weight data. For example, it is preferable to distinguish them using symbols that are not normally used as substance names, such as ":". Also, in STEP 1, if multiple substance name data and parts-of-weight data are input, each substance name data and parts-of-weight data may be confused. Therefore, it is preferable that the input be in a manner that prevents confusion between each substance name data and parts-of-weight data. For example, it is preferable to distinguish them using symbols that are not normally used as substance names, such as ",", and are different from the symbols used to distinguish between substance name data and parts-of-weight data. Furthermore, if the trained model is based on Transformer, the control unit 10 may recognize that a real number is coming by inputting the symbol <NUM> before the value (real number), and then input a real number in the next data, and input the symbol < / NUM> when the input of real numbers is finished.

[0039] In STEP 1, as a specific example, the substance name data and parts by weight data are entered as follows: [Aluminum: 50, Iron: 40, Magnesium: 10]. Additionally, arbitrary symbols to indicate the beginning (e.g., "[") and end (e.g., "[]") of the input data may be entered.

[0040] Following STEP 1, the Embedding Processing Unit 11 of the material design device 1 performs Embedding processing on the input material name data (Figure 2 / STEP 2). In the Embedding processing, the material name data is first assigned an integer unique to the material name or the label used to classify the material name. Then, the vector assigned to that integer is generated as the material name vector. That is, if the data input is [aluminum: 50, iron: 40, magnesium: 10], "aluminum" is converted to the vector X1, iron to the vector X2, and magnesium to the vector X3.

[0041] Following STEP 2, the first Linear processing unit 12 of the material design device 1 performs the first Linear processing on the input weight portion data (Figure 2 / STEP 3). In the first Linear processing, the weight portion data, which is a real number, is multiplied by a matrix, and a weight portion vector is generated as a result of the multiplication. Here, a matrix is ​​a matrix that transforms a certain value into a set that has a probability distribution representing that value. That is, if the data input is [aluminum: 50, iron: 40, magnesium: 10], "50" is transformed into the vector Y50, "40" is transformed into the vector Y40, and "10" is transformed into the vector Y10.

[0042] Here, the number of dimensions of the material name vector and the number of dimensions of the weight vector are the same. That is, the X1 vector, X2 vector, X3 vector, Y50 vector, Y40 vector, and Y10 vector are all vectors composed of the same dimensions.

[0043] In this example, STEP 3 was executed after STEP 2. However, the present invention is not limited to this order; STEP 3 may be executed after STEP 1, and STEP 2 may be executed after STEP 3. Alternatively, STEP 2 and STEP 3 may be executed in parallel after STEP 1.

[0044] Next, in STEP 4, the joint part 13 combines the substance name vector and the parts by weight vector (STEP 4). That is, for each substance, a first joint vector is generated such that each element of the substance name vector and each element of the parts by weight vector become elements. That is, for the input data of "[aluminum: 50]", a first joint vector of "[[X1], [Y50]]" is generated.

[0045] As described above, for input data composed of a plurality of substances such as "[aluminum: 50, iron: 40, magnesium: 10]", first joint vectors corresponding to each substance are generated. Then, the plurality of first joint vectors are further combined to form a first joint vector such that "[[[X1], [Y50]], [[X2], [Y40]], [[X3], [Y10]]]".

[0046] Next, in STEP 5, the generation unit 14 inputs the first joint vector into the learned model to generate a second joint vector (STEP 5). Here, the learned model is the model as described above. The learned model is a model that uses the first joint vector as an explanatory variable and the second joint vector as an objective variable.

[0047] The second joint vector is a vector obtained by superimposing the characteristic name vector of the characteristic and the characteristic quantity vector representing each amount of the characteristic. The second joint vector may be a vector representing a plurality of characteristics.

[0048] That is, in STEP 4, when the generation unit 14 is given the first combination vector [[[X1], [Y50]], [[X2], [Y40]], [[X3], [Y10]]], it generates a second combination vector that becomes [[[A1], [B1]], [[A2], [B2]]]. Here, the A1 vector is a vector representing the characteristic name of "wear amount under certain conditions" (hereinafter referred to as "characteristic name vector"), and the B1 vector is a vector representing the quantitative value of "wear amount under certain conditions" (hereinafter referred to as "characteristic quantity vector"). Similarly, the A2 vector is a vector representing the characteristic name of "friction coefficient", and the B2 vector is a vector representing the quantitative value of "friction coefficient". Note that the above characteristic names are specific examples, and as characteristics, wear amount, temperature, wear amount of the mating shaft, and test time may be used. In addition, when conducting experiments, any information useful to the experimenter may be used as the characteristic name.

[0049] The dimensionality of the second combination vector may be different from or the same as the dimensionality of the first combination vector. Also, although it is preferable that the dimensionalities of the characteristic name vector and the characteristic quantity vector representing any characteristic are the same, they may be different.

[0050] Next, in STEP 5, the separation unit 15 separates the second combination vector into a characteristic name vector and a characteristic quantity vector (STEP 6). The second combination vector that becomes [[A1], [B1]] is separated into the A1 vector and the B1 vector in the separation unit 15. When the second combination vector is a vector such as [[[A1], [B1]], [[A2], [B2]]] representing a plurality of characteristics, in the separation unit 15 or the generation unit 14, after being separated into [[A1], [B1]] and [[A2], [B2]], the characteristic name vector and the characteristic quantity vector representing the plurality of characteristics are separated.

[0051] Following STEP 6, the classification processing unit 16 performs a classification process on the characteristic name vector (STEP 7). The classification process converts the unique vectors attached to the characteristic name vector into words (tokens) representing characteristic names. For example, vector A1 is converted into the word (token) "amount of wear under certain conditions," and vector A2 is converted into the word (token) "coefficient of friction." If multiple characteristic name vectors exist, the same process is performed for each characteristic name vector.

[0052] Following STEP 7, the second Linear processing unit 17 performs the second Linear processing on the characteristic quantity vector (STEP 8). In the second Linear processing, the characteristic name vector is multiplied by a matrix, and the resulting real number characteristic quantity is generated. For example, the B1 vector is converted to the real number b1, and the B2 vector is converted to the real number b2. If there are multiple characteristic quantity vectors, the same processing is performed on each characteristic quantity vector.

[0053] In this example, STEP 8 was executed after STEP 7. However, the present invention is not limited to this order; STEP 8 may be executed after STEP 6, and STEP 7 may be executed after STEP 8. Alternatively, STEP 7 and STEP 8 may be executed in parallel after STEP 6.

[0054] Following STEP 8, the output unit 40 outputs the obtained characteristic name and characteristic quantity (STEP 9). That is, if the generation unit 14 generates a second bond vector such that [[[A1][B1]][[A2][B2]]], then information such as [Amount of wear under certain conditions: b1, coefficient of friction: b2] is output to the output unit 40. At this time, the token generated in STEP 7 is converted into specific character information. The generation unit 14 may also output any symbols to indicate the beginning (e.g., "[") and end (e.g., "[]") of the output data.

[0055] In STEP 9, characteristic name data and characteristic quantity data may be confused. Therefore, it is preferable that the output be in a manner that prevents confusion between characteristic name data and characteristic quantity data. For example, it is preferable to distinguish them using symbols that are not normally used as characteristic names, such as ":". Also, in STEP 9, if multiple characteristic name data and characteristic quantity data are output, each characteristic name data and characteristic quantity data may be confused. Therefore, it is preferable that the output be in a manner that prevents confusion between each characteristic name data and characteristic quantity data. For example, it is preferable to distinguish them using symbols that are not normally used as characteristic names, such as ",", and symbols different from the symbols used to distinguish between characteristic name data and characteristic quantity data. Furthermore, if the trained model is based on Transformer, the value (real number) may be output with the symbols <NUM> and < / NUM> placed between them.

[0056] The material design apparatus 1 with the above method and configuration can estimate the properties of a material containing multiple substances.

[0057] Furthermore, in the above embodiment, a material design apparatus 1 for estimating the properties of a material containing multiple substances was described. However, the present invention is not limited to this, and another embodiment may provide a material design apparatus 1 for estimating the properties of a material having an arbitrary material composition under arbitrary measurement conditions.

[0058] In this configuration of material design apparatus 1, information such as the material name representing the type of material, the parts by weight representing its composition ratio, and test conditions (e.g., whether it is a thrust test or a journal test) or bearing type conditions (whether it is bearing A or bearing B) are input via the input unit 30 as input data, and characteristic name data representing the type of characteristic and characteristic quantity data representing the quantitative value of the characteristic are output via the output unit 40 as output data. The other configurations are substantially the same as the configuration described earlier, so details are omitted.

[0059] Furthermore, in another embodiment of the present invention, a material design apparatus 1 for estimating the properties of a material having an arbitrary material composition under specific measurement conditions may be provided.

[0060] In this configuration of material design apparatus 1, the material name representing the type of substance and its weight representing its constituent ratio are input as input data via the input unit 30, and the characteristic name data representing the type of characteristic and characteristic quantity data representing the quantitative value of the characteristic under specific test conditions are output via the output unit 40 as output data. The other configurations are substantially the same as the configuration described earlier, so the details are omitted.

[0061] Furthermore, in the above embodiment, a material design device 1 for estimating the properties of a material containing multiple substances was described. However, the present invention is not limited to this, and another embodiment provides a material design device 1 for estimating the material composition of a material having arbitrary properties.

[0062] In this configuration of material design apparatus 1, the input data consists of characteristic name data representing the type of characteristic and characteristic quantity data representing the quantitative value of the characteristic, which are input via the input unit 30. The output data consists of the substance name representing the type of substance and its weight portion representing its constituent ratio, which are output via the output unit 40. The other configurations are substantially the same as those described above, so details are omitted. In other words, in this configuration as well, the test conditions may be input as input data, or the test conditions may be output as output data.

[0063] The material design apparatus 1 with the above method and configuration can estimate the material composition of a material having certain properties.

[0064] Next, embodiments of the present invention will be described.

[0065] (First Embodiment) The first embodiment is a model in which the trained model is an encoder-decoder architecture having an encoder and a decoder. The first embodiment is a material design device 1 that takes component A and component B and their respective weight parts as input data (N=3) and outputs wear amount and friction coefficient.

[0066] Figure 3 shows the learning curve of the material design apparatus 1 according to the first embodiment. In Figure 3, the horizontal axis represents the number of epochs, and the vertical axis represents the learning error. Referring to Figure 3, it can be seen that the loss value (vertical axis) approaches 0 at a convergence calculation of 10,000 epochs, suggesting that learning is possible with such data.

[0067] Table 1 shows the inference results of the material design device 1. Table 1 also shows the characteristics of the training data for the material composition of the inference data. Note that the wear amount and friction coefficient of the inference data are under the requirements of the training data. Here, the friction coefficient has three types: "friction coefficient HIGH", "friction coefficient MID", and "friction coefficient LOW". "Friction coefficient HIGH", "friction coefficient MID", and "friction coefficient LOW" are the kinetic friction coefficient at high speed, kinetic friction coefficient at medium speed, and kinetic friction coefficient at low speed, respectively.

[0068]

[0069] Referring to Table 1, some values ​​are similar for the friction coefficient HIGH, but the inferred value and the input value differ by a factor of 10. However, it became clear that the error in the wear amount is small, not only in terms of the order of magnitude of the numerical values, but also in the numerical values ​​themselves. It also became clear that the error in the inferred value is large depending on the order of magnitude of the numerical values.

[0070] The characteristics with large errors are thought to be due to overfitting, resulting from the small number of data points (N is 3 in Example 1) and the large number of epochs (10,000). However, it became clear that accurate inference data was obtained for some characteristics.

[0071] (Second Embodiment) The second embodiment is a model in which the trained model is an encoder-decoder architecture having an encoder and a decoder. The second embodiment is a material design device 1 that takes component A, component B and other compounds and their weight parts as input data (N=7) and outputs wear amount and friction coefficient.

[0072] Figure 4 shows the learning curve of the material design apparatus 1 according to the second embodiment. In Figure 4, the horizontal axis represents the number of epochs, and the vertical axis represents the learning error. In the second embodiment, the number of epochs was set to 5000 because, in the first embodiment, there was almost no change in the learning error value beyond approximately 5000 epochs. Referring to Figure 4, in the convergence calculation with 5000 epochs, although there are some inflection points along the way, the learning error value decreases as the number of epochs increases, suggesting that learning is progressing.

[0073] Table 2 shows the inference results of the material design apparatus 1 in the second embodiment. Table 2 also shows the characteristics of the training data for the material composition of the inference data. Note that the wear amount and friction coefficient of the inference data are based on the requirements of the training data.

[0074]

[0075] Similar to the first embodiment, it was found that the amount of wear was approximately equal between the training data and the inference results. Furthermore, for relatively large friction coefficients (to two decimal places), such as friction coefficients MID and LOW, there was a tendency for the order of magnitude of the values ​​to be similar. This is presumed to be a result of calculation errors in the sum-of-products calculation caused by matrix calculations.

[0076] (Third Embodiment) The third embodiment is a model in which the trained model is an encoder-decoder architecture having an encoder and a decoder. The third embodiment is a material design device 1 that takes component A, component B and other compounds and their weight parts as input data (N = 10⁹) and outputs wear amount and friction coefficient.

[0077] Figure 5 shows the learning curve of the material design apparatus 1 according to the third embodiment. In Figure 5, the horizontal axis represents the number of epochs, and the vertical axis represents the learning error. In the third embodiment, the number of epochs is set to 20,000, which is due to the larger amount of input data (N) compared to the first and second embodiments. Referring to Figure 5, the learning error value decreases with increasing number of epochs, suggesting that learning is taking place. However, it is clear that the decrease in the learning error value is slower compared to Figures 3 and 4. This is thought to affect the accuracy of the calculation results in the third embodiment.

[0078] Figure 6 shows a comparison of the inference results and training data of the material design apparatus 1 in the third embodiment. Referring to Figure 6, except for the friction coefficient HIGH, Exact (training data) generally has a positive correlation with Pred (inference results), indicating that learning has been successful. In particular, for wear amount, a log-log graph was used because the range of values ​​is wide, but even with a wide range of values, the inferred values ​​and training values ​​tended to be in close agreement, indicating that learning was successful. On the other hand, for the friction coefficient HIGH, as shown in the circled area in the figure, there were many results inferring negative values, which suggests that learning was not successful. Furthermore, for the friction coefficient MID and friction coefficient LOW, which take decimal values, the inferred values ​​and training values ​​generally have a positive correlation, suggesting that learning has been successful despite some variability.

[0079] Comparing the well-learned wear amount with the relatively well-learned friction coefficient MID and friction coefficient LOW, the wear amount is two digits in real numbers, and the friction coefficients are two digits in real numbers. This trend shows that the number of digits in the friction coefficient HIGH value (less than three digits) and the number of digits in the wear amount differ by up to five digits (a difference of more than 10,000), suggesting that the number of digits to be learned may be too large. This is thought to be because it cannot be handled within the limits of the computer's internal calculation processing (the numerical data type used here is float32). The characteristics of the learning data in the material composition of the inference data are also shown. From these results, it is suggested that the wear amount is reliably learned, and it became clear that even among the friction coefficients, the trend itself can be captured if it is a relatively large value.

[0080] (Fourth Embodiment) The fourth embodiment is a model in which the trained model is a recurrent neural network that has undergone transfer learning. The fourth embodiment is a material design device 1 that takes specific materials and their weights as input data and outputs the amount of wear, friction coefficient, and temperature (temperature generated by frictional heat after the experiment). As for transfer learning, two methods are used: using the results learned with data group A (N=505) for data group B (N=288) (Fourth-First Embodiment), and using the results learned with data group B for data group A (Fourth-Second Embodiment).

[0081] Figure 7 shows the learning curve for data group A of the material design apparatus 1 according to the fourth embodiment. The horizontal axis of Figure 7 is the number of epochs, and the vertical axis is the learning error. In the fourth embodiment, the number of epochs for the learning curve for data group A is set to 10,000,000. Data group A is a data group related to material composition and material properties (amount of wear, coefficient of friction, temperature).

[0082] Figure 8 shows the learning curve for data group B of the material design apparatus 1 according to the fourth embodiment. The horizontal axis of Figure 8 represents the number of epochs, and the vertical axis represents the learning error. In the fourth embodiment, the number of epochs for the learning curve for data group B is set to 20,000. Data group B is a data group relating to material composition and material properties (coefficient of friction).

[0083] Figures 7 and 8 clearly show that learning progresses as the number of epochs increases. In Figure 7, although the horizontal axis is linear, the tendency for the learning error to decrease sharply at the points indicated by the arrows suggests that a similar trend is observed for both data set A and data set B.

[0084] Figures 9 and 10 show the learning curves for the results of transfer learning on data set B after learning on data set A (Example 4-1) and the results of transfer learning on data set A after learning on data set B (Example 4-2), respectively.

[0085] Figures 9 and 10 clearly show that learning can be performed even when transfer learning is carried out between different data sets. Referring particularly to Figure 9, it was found that in the 4-1 embodiment, the learning error was small from the early stages of learning and did not change much even at the end of learning. Referring also to Figure 10, it was found that in the 4-2 embodiment, the learning error was large at first but gradually decreased with increasing epochs, and at the end of learning, the loss value was smaller than in the 4-1 embodiment. This is thought to be because data set A has a large number of responses (amount of wear, coefficient of friction, temperature), while data set B has a small number of responses (coefficient of friction only), and the amount of learning required for the material design device 1 using data set A is large. Figure 10 shows the opposite, suggesting that the change in learning error was small because only the coefficient of friction needed to be used as a response.

[0086] Figure 11 shows a comparison between the inference results of the material design device 1, which was trained based on data set A, and the training data. Referring to Figure 11, similar to the first to third embodiments, the inference results and the training data are generally consistent, suggesting that the device was trained accurately. Similarly, Figure 12 shows a comparison between the inference results of the material design device 1, which was trained based on data set B, and the training data. Referring to Figure 12, similar to Figure 11, the inference results and the training data are generally consistent, suggesting that the device was trained accurately.

[0087] Figure 13 shows a comparison between the inference results of the material design device 1 and the training data in the 4-1 embodiment (material design device 1 that performs transfer learning of data group A based on data group B). Referring to Figure 13, although there is considerable variability, the inference results generally show agreement with the training data, making it clear that learning and inference are possible. From this, it is suggested that the learning model for data group A is similar in concept to that of data group B (in terms of material, quantity, and response (characteristic value)). Comparing Figure 13, which is the result of transfer learning, with Figure 12, which is the result of normal learning, it was found that although there are some differences, the inference is comparable. Therefore, it is suggested that even when using transfer learning, the inference will be almost the same as when using normal learning.

[0088] Figure 14 shows a comparison of the inference results of the material design device 1 and the training data in the 4-2 embodiment (material design device 1 that performs transfer learning of data group B based on data group A). ​​Referring to Figure 14, the amount of wear, friction coefficient, and temperature all showed similar trends to those shown in Figure 11. This suggests that the approach is similar to that in Figure 11 in terms of the response (characteristics) to the material.

[0089] The results in Figures 13 and 14 suggest that learning is possible using transfer learning. As a result, it is suggested that learning is generally possible even when the data is standardized.

[0090] (Fifth Embodiment) In the fifth embodiment, the trained model is a model based on a recurrent neural network. Furthermore, the fifth embodiment is a material design apparatus 1 that takes specific materials and their weights, as well as information identifying whether it is bearing A or bearing B, as input data (N=287), and outputs wear amount, friction coefficient, and temperature. Here, bearing A and bearing B are defined by the type of bearing.

[0091] Figure 15 shows the learning curve for the material design apparatus 1 according to the fifth embodiment ("material design apparatus 1 with information to identify bearing types input"). Figure 16 shows the learning curve for the material design apparatus 1 in the fifth embodiment when information to identify bearing types is not input. In Figures 15 and 16, the horizontal axis represents the number of epochs, and the vertical axis represents the learning error.

[0092] Figure 17 shows a comparison of the inference results of the material design device 1 according to the fifth embodiment with the training data. Figure 18 shows a comparison of the inference results of the material design device 1 with the training data in the fifth embodiment when no information identifying the type of bearing is input.

[0093] Referring to Figures 15 and 16, it became clear that the learning process differed little regardless of whether data was identified or not. Furthermore, referring to Figures 17 and 18, it was suggested that the presence or absence of data identification did not have a significant impact on the overall inference, but differences were observed in the areas circled in Figures 17 and 18. It is suggested that these differences in the circled areas are likely due to the influence of the initial random numbers used in the learning weights. These errors are considered to be the so-called "statistical" part and can be reduced by performing multiple learning processes and selecting the appropriate model from among them.

[0094] From these results, it was suggested that in the fifth embodiment, the general distribution of inferred values ​​followed the same trend regardless of the presence or absence of identification information, and rather, the influence of the initial random numbers for the learning weights had a greater impact.

[0095] (Sixth Embodiment) In the sixth embodiment, the trained model is a transformer-based model. The sixth embodiment is a material design device 1 that takes specific materials and their weight as input data (N=500) and outputs wear amount, friction coefficient, and temperature. Experimental parameters from journal tests are also included as input data.

[0096] Figure 19 shows the learning curve for the material design apparatus 1 according to the sixth embodiment. Figure 20 shows the verification curve for the material design apparatus 1 according to the sixth embodiment. The verification curve represents the error calculated using data not used for learning (in this case, learning and testing are rotated randomly). In Figures 19 and 20, the horizontal axis is the number of epochs, and the vertical axis is the learning error. In this learning process, the maximum number of epochs is 8192.

[0097] In Figure 19, the darker circles represent the error in the regression portion (loss of real numbers), and the lighter circles represent the error in the classification portion (loss of words). From this, it can be seen that the error in the classification portion decreases logarithmically. The minimum value of the error in the classification portion is around 0.00001, and it decreases with increasing epochs, clearly indicating that learning is being carried out reliably. On the other hand, although the error curve for the regression portion is logarithmic, the slope of decrease is small, and even at 8192 epochs, the decrease remains at around 0.1.

[0098] The validation curve in Figure 20 is the sum of the error in the regression part and the error in the classification part. Therefore, the validation curve in Figure 20 shows a trend that is generally similar to the learning curve, with the final epoch being around 0.1. The error in the regression part is calculated as L2 (error squared), so if the error is less than 1, the error will be a small value. However, since the error is around 0.1, it is suggested that the inference error in the regression results is large.

[0099] Figure 21 shows a comparison of the inference results of the material design apparatus 1 according to the fifth embodiment with the training data. In Figure 21, the number of data points for which complete inference was possible for N=500 was 436. This was due to the values ​​diverging during the inference process. Referring to Figure 21, all wear amounts were inferred, and only a very small number of temperature values ​​could be inferred. It became clear that the coefficient of friction tended to be inferred to be slightly smaller, the wear amount to be slightly larger, and the temperature tended to be inferred to be larger as the temperature increased.

[0100] When comparing Figures 17 and 18, which show the inference results of the fifth embodiment, Figure 21 shows results that are closer to the correct answer, although the sample size (N) is different. When comparing the variability of the data, it was suggested that variability is expressed in both the fifth and sixth embodiments. Focusing on the area enclosed by the ellipse for wear amount, the variability trend is almost the same in the fifth and sixth embodiments. Focusing on the friction coefficient, the variability trend is also generally similar in the area enclosed by the ellipse. This trend suggests that a regression model can be formed even in the sixth embodiment, which is a transformer-based model.

[0101] Figures 22 and 23 show the estimated wear amount and friction coefficient when a material containing 20 parts by weight of component A and 2.5 parts by weight of component D17 relative to the total weight estimated in the sixth embodiment was tested using a journal test method, with a clearance of 55.66667 μm, a speed of 2 m / s, and a bearing type of bearing A. Figure 22(a) shows the wear amount when the amount of component A is varied from the basic conditions, Figure 22(b) shows the friction coefficient when the amount of component A is varied from the basic conditions, Figure 22(c) shows the wear amount when the amount of component D is varied from the basic conditions, and Figure 22(d) shows the friction coefficient when the amount of component D is varied from the basic conditions.

[0102] Referring to Figures 22(a) and 22(b), a step-like change occurs in the amount of wear and the coefficient of friction when the weight of component A increases from approximately 10 parts by weight to 15 parts by weight. This is thought to be due to the fact that the learning data for the sixth embodiment only includes data for component A at 10 and 20 parts by weight, resulting in a data bias. Furthermore, focusing on Figures 22(c) and 22(d), a step-like change occurs in the weight of component D at around 2.5 parts by weight, 5 parts by weight, 10 parts by weight, and 15 parts by weight, similar to when the weight of component A is changed. Even within the step-like changes, there are slight fluctuations near each value, and behavior that seems to be learned from other data is observed. Therefore, it is thought that the learning accuracy will improve by collecting further experimental results.

[0103] Figure 23 shows the coefficient of friction when the journal test speed is varied from the baseline conditions. In Figure 23, the triangles represent experimental data, and the circles represent the inference results. It can be seen that the circles, which are the inference results, are roughly interpolated between the triangles, which are the experimental data. The inference value for a speed of 0.3 m / s, which lies between speeds of 0.25 m / s and 0.4 m / s, is slightly smaller than the experimental value, but the trend suggests that interpolation is occurring. Looking at other speeds, it was also confirmed that the inference value for a speed of 0.5 m / s, which lies between speeds of 0.4 m / s and 0.6 m / s, is roughly interpolative. Looking at the overall trend, for speeds of 0.1 m / s and above, the trends are generally similar, suggesting that numerical values ​​can be interpolated and that the trend itself can be roughly learned.

[0104] Similarly, Figures 24 and 25 show the estimated wear amount and friction coefficient when a material containing 20 parts by weight of component A, 7 parts by weight of component C, 7 parts by weight of component D, and 3 parts by weight of component B relative to the total weight estimated in the sixth embodiment was tested using a thrust test method at a speed of 5 m / s and bearing type A. Figure 24(a) shows the wear amount when the amount of component B is varied from the basic conditions, Figure 24(b) shows the friction coefficient when the amount of component B is varied from the basic conditions, Figure 24(c) shows the wear amount when the amount of component C is varied from the basic conditions, Figure 24(d) shows the friction coefficient when the amount of component C is varied from the basic conditions, Figure 24(e) shows the wear amount when the amount of component D is varied from the basic conditions, and Figure 24(f) shows the friction coefficient when the amount of component D is varied from the basic conditions. In Figures 24(a), 24(b), and 24(d), there are areas where no plots are present on the horizontal axis. These areas represent regions where inference could not be made.

[0105] Referring to Figures 24(a) and 24(b), it was confirmed that there is an increase or decrease in the response value depending on the amount of material. In Figure 24(a), a behavior in which the maximum value is reached with the amount of wear is shown, while in Figure 24(b) shows a situation in which there are few inference results and inference is not possible. Thus, it is suggested that there are cases in which the response can only be inferred near the learned value.

[0106] Referring to Figures 24(c) and 24(d), it was confirmed that while the amount of wear could be inferred across almost the entire range of varying amounts of component C, the coefficient of friction could not be inferred in all cases. It was confirmed that the amount of wear was minimal at 5 parts by weight and 10 parts by weight of component C. Furthermore, it was confirmed that the coefficient of friction was minimal at 5 parts by weight of component C, and extremely large at 11 parts by weight of component C. This suggests that although learning has been achieved, there are problems with the inference accuracy.

[0107] Referring to Figures 24(e) and 24(f), the change in wear amount due to the amount of component D (Figure 24(e)) showed a trend almost identical to the change in wear amount due to the amount of component C (Figure 24(c)). The change in the coefficient of friction due to the amount of component D (Figure 24(f)) was confirmed to be significant when the weight of component D was between 5 and 15 parts by weight.

[0108] Based on these results, although learning is possible, it can be inferred that the response behavior of bearing A is not necessarily extrapolative when the amount of material is varied. This is likely because there is a bias in the amount of material used as training data.

[0109] Figure 25(a) shows the coefficient of friction when the thrust test speed is varied from the basic conditions. Figure 25(b) shows the amount of wear when the thrust test speed is varied from the basic conditions.

[0110] Referring to Figure 25, it was confirmed that there are parts of the relationship between speed and friction coefficient, and speed and wear, particularly regarding speed, where inference is not possible. Referring to Figure 25(a), the inference result was obtained that the friction coefficient is high at low speeds and decreases as the speed increases. Referring to Figure 25(b), although the range is limited, the relationship between speed and wear also showed that the wear was 2.5 μm at a speed of 0.01 m / s, and the wear tended to decrease with increasing speed. This relationship suggests that the high friction coefficient indicates the "solid lubrication region," and the decreasing friction coefficient indicates the "mixed lubrication region." Wear also tends to be large in the "solid lubrication region" and gradually decrease in the "mixed lubrication region," suggesting that the lubrication region has generally been learned.

[0111] Figure 26 shows the estimated wear amount and friction coefficient when a material containing 20 parts by weight of component A, 3 parts by weight of component E, and 3 parts by weight of component D relative to the total weight estimated in the sixth embodiment was subjected to a test using a thrust test method, a speed of 3 m / s, and bearing type A.

[0112] Figure 26(a) shows the coefficient of friction when the thrust test speed is varied from the base conditions. Figure 26(b) shows the amount of wear when the thrust test speed is varied from the base conditions.

[0113] The inference results in Figure 26 were completely different from those in Figure 25. Referring to Figure 26, regarding the relationship between speed and friction coefficient, it was confirmed that there is a cluster around a speed of 0.1 m / s, and below that speed, the friction coefficient fluctuates between 0.06 and 0.07. Furthermore, as the speed increases, above 0.5 m / s, the friction coefficient remains below 0.01, suggesting that while the speed range has been learned, the lubrication range has not yet been learned. On the other hand, the relationship between speed and wear amount also showed a slight decrease up to 0.1 m / s, but above that speed, the behavior fluctuated up and down around 2.45 μm, confirming that extrapolation was not possible.

[0114] From these results, it was confirmed that extrapolation can be performed successfully in some cases and not in others, depending on the data being inferred. This is likely because extrapolation is not successful when there is insufficient data on the marginal distribution, not just on a single data point.

[0115] Figure 27 shows the estimated friction coefficient of a bearing when the bearing type is bearing B, using a material containing 100 parts by weight of component F, 100 parts by weight of component G, 125 parts by weight of component H, 100 parts by weight of component I, and 3 parts by weight of component J relative to the total weight estimated in the sixth embodiment. Note that component H is estimated after being input separately as component H1, component H2, and component H3. Figure 27(a) shows the coefficient of friction when the amount of component H1 is varied from the basic conditions, Figure 27(b) shows the coefficient of friction when the amount of component I is varied from the basic conditions using component H1, Figure 27(c) shows the coefficient of friction when the amount of component H2 is varied from the basic conditions using component H2, Figure 27(d) shows the coefficient of friction when the amount of component I is varied from the basic conditions using component H2, Figure 27(e) shows the coefficient of friction when the amount of component H3 is varied from the basic conditions using component H3, and Figure 27(f) shows the coefficient of friction when the amount of component I is varied from the basic conditions using component H3. Note that when component I is varied, the amount of all components H is 125 parts by weight.

[0116] Referring to Figures 27(a), 27(c), and 27(e), a comparison of experimental data and inference results shows that, although there is some variability, the inference generally falls within the central range. This is thought to be due to the small amount of experimental data, which prevents the demonstration of variability. However, since other similar experimental results have also been learned, it is thought that the result is a generally average curve. On the other hand, referring to Figures 27(b), 27(d), and 27(f), it was observed that the coefficient of friction fluctuates when the amount of component I is small, but tends to become almost constant once the amount of component I exceeds about 40 parts by weight. When the amount is small, the coefficient of friction fluctuates based on the relative amount with component H, but once it reaches a certain amount (proportion of the whole), it is shown that the material is almost independent of the coefficient of friction. This indicates that the amount of component I does not have much effect on the coefficient of friction, and it can be inferred that this phenomenon has been largely inferred.

[0117] Figure 28 shows the estimated coefficient of friction for a bearing of type B, using a material containing 100 parts by weight of component F, 100 parts by weight of component G, 100 parts by weight of component K, 1 part by weight of component L, and 100 parts by weight of component I, relative to the total weight estimated in the sixth embodiment. Figure 28(a) shows the coefficient of friction when the amount of component L is varied from the basic conditions, and Figure 28(b) shows the coefficient of friction when the amount of component I is varied from the basic conditions.

[0118] Referring to Figure 28, when the amount of component L was varied, the coefficient of friction increased from 2 parts by weight to about 10 parts by weight of component L, and the coefficient of friction remained almost constant even when the amount was increased beyond that point. In fact, increasing the amount of graphite generally increases the coefficient of friction, but this result is thought to be due to the fact that there was no result in the peripheral distribution when the amount was increased logarithmically like this. On the other hand, when the amount of component I was varied, there was one point where the coefficient of friction decreased, but it was confirmed that there was almost no change. This result is thought to be due to the fact that component I is input in almost all of the data.

[0119] In summary, our investigation into bearing B revealed that some materials yielded results similar to those predicted by empirical rules, while others did not. Furthermore, it was suggested that even with a small number of individual experiments, if similar results were observed collectively, it is possible to make general inferences about certain materials.

[0120] Although one embodiment of the present invention and its examples have been described above, it is clear that the present invention is not limited to the above embodiments or examples, and that the scope of the present invention can be modified and altered to the extent that is obvious to those skilled in the art. It is also clear that the scope of the present invention is not limited to the above embodiments or examples, but also includes modified and altered versions thereof.

[0121] 1...Material design device 1, 10...Control unit, 11...Embedding processing unit, 12...First Linear processing unit, 13...Coupling unit, 14...Generation unit, 15...Separation unit, 16...Classification processing unit, 17...Second Linear processing unit, 20...Storage unit, 30...Input unit, 40...Output unit.

Claims

1. A material design apparatus for designing a material containing multiple substances, comprising: an input unit for inputting the substance names of the multiple substances and the weight portion, which is the amount of the substance corresponding to the substance name relative to the total mass; an embedding processing unit for which embedding processing is performed for the multiple substance names and generates a substance name vector; a first linear processing unit for which linear processing is performed for the weight portion corresponding to the multiple substance names and generates a weight portion vector; a coupling unit for which the substance vector and weight portion vector corresponding to the multiple substances are superimposed to generate a first coupling vector; a generation unit for which the first coupling vector is input to a trained model and generates a second coupling vector by superimposing a characteristic name vector representing the characteristic name of at least one characteristic of the material to be designed and a characteristic quantity vector representing the quantity of at least one characteristic; a separation unit for separating the superposition of the second coupling vector; a classification processing unit for which classification processing is performed for at least one characteristic name vector and generates a characteristic name; and a second linear processing unit for which linear processing is performed for at least one characteristic quantity vector and generates a characteristic quantity. A material design apparatus characterized by comprising an output unit that outputs the characteristic name and characteristic quantity of at least one or more characteristics.

2. A material design apparatus for designing a material containing multiple substances, comprising: an input unit for inputting characteristic names and quantities corresponding to at least one characteristic; an embedding processing unit for performing embedding processing on at least one characteristic name and generating a characteristic name vector; a first linear processing unit for performing linear processing on quantities corresponding to at least one characteristic name and generating a characteristic quantity vector; a coupling unit for superimposing the characteristic name vector and characteristic quantity vector corresponding to at least one characteristic and generating a first coupling vector; a generation unit for inputting the first coupling vector into a trained model and generating a second coupling vector by superimposing a substance name vector representing the substance names of the multiple substances constituting the material to be designed and a weight portion vector representing the weight portion of the multiple substances as a whole; a separation unit for separating the superposition of the second coupling vector; a classification processing unit for performing classification processing on multiple substance name vectors and generating substance names; and a second linear processing unit for performing linear processing on multiple weight portion vectors and generating weight portions. A material design apparatus characterized by comprising an output unit that outputs the names and weights of multiple substances.

3. A material design apparatus according to claim 1 or 2, characterized in that the trained model is a model based on an encoder-decoder architecture having an encoder and a decoder, a recurrent neural network (RNN) model, or a transformer model.

4. A material design method for designing a material containing multiple substances, comprising: an input step of inputting the substance names of the multiple substances and the parts by weight which are the amounts of the substances corresponding to the substance names relative to the total mass; an embedding step of performing embedding processing on the multiple substance names to generate a substance name vector; a first linear processing step of performing linear processing on the parts by weight corresponding to the multiple substance names to generate a parts by weight vector; a coupling step of superimposing the substance vectors and parts by weight vectors corresponding to the multiple substances to generate a first coupled vector; a generation step of inputting the first coupled vector into a trained model to generate a second coupled vector which is a superimposed characteristic name vector representing the characteristic name of at least one or more properties of the material to be designed and a characteristic quantity vector representing the quantity of at least one or more properties; a separation step of separating the superposition of the second coupled vector; a classification processing step of performing classification processing on at least one or more characteristic name vectors to generate characteristic names; and a second linear processing step of performing linear processing on at least one or more characteristic quantity vectors to generate characteristic quantities. A material design method characterized by comprising an output step that outputs the characteristic name and characteristic quantity of at least one or more characteristics.

5. A material design method for designing a material containing multiple substances, comprising: an input step of inputting a characteristic name and a quantity corresponding to at least one characteristic; an embedding step of performing embedding processing for at least one characteristic name to generate a characteristic name vector; a first linear processing step of performing linear processing for the quantity corresponding to at least one characteristic name to generate a characteristic quantity vector; a coupling step of superimposing the characteristic name vector and the characteristic quantity vector corresponding to at least one characteristic to generate a first coupled vector; a generation step of inputting the first coupled vector into a trained model to generate a second coupled vector by superimposing a substance name vector representing the substance name of each of the multiple substances constituting the material to be designed and a weight portion vector representing the weight portion of the multiple substances in total; a separation step of separating the superposition of the second coupled vector; a classification processing step of performing classification processing on the multiple substance name vectors to generate substance names; and a second linear processing step of performing linear processing on each of the multiple weight portion vectors to generate their respective weight portions. A material design method characterized by comprising an output step that outputs the names of multiple substances and the corresponding parts by weight.

Images